From: Wolfgang Bangerth Date: Fri, 7 Jan 2022 19:52:12 +0000 (-0700) Subject: Mention Kahan summation and discuss 'double' vs 'long double'. X-Git-Tag: v9.4.0-rc1~643^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=8962d46ab8a1421d0c08fe8d74a25c669311decf;p=dealii.git Mention Kahan summation and discuss 'double' vs 'long double'. --- diff --git a/examples/step-10/doc/results.dox b/examples/step-10/doc/results.dox index 628bfc69d2..8b6e7a0471 100644 --- a/examples/step-10/doc/results.dox +++ b/examples/step-10/doc/results.dox @@ -190,7 +190,7 @@ cells eval.pi error 320 3.1415926535897576 3.5527e-14 7.94 1280 3.1415926535897896 3.5527e-15 3.32 5120 3.1415926535897940 8.8818e-16 2.00 - @endcode +@endcode @note Once the error reaches a level on the order of $10^{-13}$ to $10^{-15}$, it is essentially dominated by @@ -198,6 +198,9 @@ cells eval.pi error in internal computations. Since these things change, the precise values and errors change from release to release at these round-off levels, though the overall order of errors should of course remain the same. + See also the comment below in the section on + Possibilities for extensions about how to compute + these results more accurately. One of the immediate observations from the output above is that in all cases the values converge quickly to the true value of @@ -223,3 +226,97 @@ convergence. (This effect is also discussed in detail in the following publication: A. Bonito, A. Demlow, and J. Owen: "A priori error estimates for finite element approximations to eigenvalues and eigenfunctions of the Laplace-Beltrami operator", submitted, 2018.) + + + +

Possibilities for extensions

+ + +As the table of numbers copied from the output of the program shows above, +it is not very difficult to compute the value of $\pi$ to 13 or 15 digits. But, +the output also shows that once we approach the level of accuracy with which +`double` precision numbers store information (namely, with roughly 16 digits +of accuracy), we no longer see the expected convergence order and the error +no longer decreases with mesh refinement as anticipated. This is because both +within this code and within the many computations that happen within deal.II +itself, each operation incurs an error on the order of $10^{-16}$; adding +such errors many times over then results in an error that may be on the +order of $10^{-14}$, which will dominate the discretization error after +a number of refinement steps and consequently destroy the convergence rate. + +The question is whether one can do anything about this. One thought is to +use a higher-precision data type. For example, one could think of declaring +both the `area` and `perimeter` variables in `compute_pi_by_area()` and +`compute_pi_by_perimeter()` with data type `long double`. `long double` +is a data type that is not well specified in the C++ standard but at least +on Intel processors has around 19, instead of around 16, digits of accuracy. +If we were to do that, we would get results that differ from the ones shown +above. However, maybe counter-intuitively, they are not uniformly better. +For example, when computing $\pi$ by the area, at the time of writing +these sentences we get these values with `double` precision for degree 4: +@code + 5 3.1415871927401144 5.4608e-06 - + 20 3.1415926314742491 2.2116e-08 7.95 + 80 3.1415926535026268 8.7166e-11 7.99 + 320 3.1415926535894005 3.9257e-13 7.79 + 1280 3.1415926535899774 1.8430e-13 1.09 + 5120 3.1415926535897669 2.6201e-14 2.81 +@endcode +On the other hand, the results are as follows when using `long double`: +@code + cells eval.pi error + 5 3.1415871927401136 5.4608e-06 - + 20 3.1415926314742446 2.2116e-08 7.95 + 80 3.1415926535026215 8.7172e-11 7.99 + 320 3.1415926535894516 3.4157e-13 8.00 + 1280 3.1415926535897918 1.5339e-15 7.80 + 5120 3.1415926535897927 5.2649e-16 1.54 +@endcode +Indeed, here we get results that are approximately 50 times as accurate. +On the other hand, when computing $\pi$ by the perimeter, we get this with +`double` precision: +@code + 5 3.1415921029432572 5.5065e-07 - + 20 3.1415926513737582 2.2160e-09 7.96 + 80 3.1415926535810699 8.7232e-12 7.99 + 320 3.1415926535897576 3.5527e-14 7.94 + 1280 3.1415926535897896 3.5527e-15 3.32 + 5120 3.1415926535897940 8.8818e-16 2.00 +@endcode +Whereas we get the following with `long double`: +@code + 5 3.1415921029432572 5.5065e-07 - + 20 3.1415926513737595 2.2160e-09 7.96 + 80 3.1415926535810703 8.7230e-12 7.99 + 320 3.1415926535897576 3.5705e-14 7.93 + 1280 3.1415926535897918 1.3785e-15 4.70 + 5120 3.1415926535897944 1.3798e-15 -0.00 +@endcode +Here, using `double` precision is more accurate by about a factor of +two. (Of course, in all cases, we have computed $\pi$ with more +accuracy than any engineer would ever want to know.) + +What explains this unpredictability? In general, round-off errors can +be thought of as random, and add up in ways that are not worth thinking +too much about; we should therefore always treat any accuracy beyond, say, +thirteen digits as suspect. Thus, it is probably not worth spending +too much time on wondering why we get different winners and losers in the +data type exchange from `double` and `long double`. The accuracy of the +results is also largely not determined by the precision of the data type +in which we accumulate each cell's (or face's) contributions, but the +accuracy of what deal.II gives us via FEValues::JxW() and FEFaceValues::JxW(), +which always uses `double` precision and which we cannot directly affect. + +But there are cases where one can do something about the precision, and it +is worth at least mentioning the name of the most well-known algorithm in +this area. Specifically, what we are doing when we add contributions into +the `area` and `perimeter` values is that we are adding together *positive* +numbers as we do here. In general, the round-off errors associated with each +of these numbers is random, and if we add up contributions of substantially +different sizes, then we will likely be dominated by the error in the largest +contributions. One can avoid this by adding up numbers sorted by their +size, and this may then result in marginally more accurate end results. +The algorithm that implements this is typically called +Kahan's summation algorithm. +While one could play with it in the current context, it is likely not going +to improve the accuracy in ways that will truly matter. diff --git a/examples/step-10/step-10.cc b/examples/step-10/step-10.cc index 88bb451ee3..237207efd9 100644 --- a/examples/step-10/step-10.cc +++ b/examples/step-10/step-10.cc @@ -349,9 +349,7 @@ namespace Step10 } // Then store the evaluated values in the table... table.add_value("eval.pi", static_cast(perimeter / 2.0)); - table.add_value("error", - static_cast( - std::fabs(perimeter / 2.0 - numbers::PI))); + table.add_value("error", std::fabs(perimeter / 2.0 - numbers::PI)); } // ...and end this function as we did in the previous one: